Cold-start Sequential Recommendation via Meta Learner

Authors: Yujia Zheng, Siyi Liu, Zekun Li, Shu Wu4706-4713

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted on three real-world datasets verify the superiority of Mecos, with the average improvement up to 99%, 91%, and 70% in HR@10 over state-of-the-art baseline methods.
Researcher Affiliation Academia Yujia Zheng,1 Siyi Liu, 1 Zekun Li, 2, 3 Shu Wu 4, 5, * 1 University of Electronic Science and Technology of China 2 School of Cyber Security, University of Chinese Academy of Sciences 3 Institute of Information Engineering, Chinese Academy of Sciences 4 School of Artificial Intelligence, University of Chinese Academy of Sciences 5 Institute of Automation and Artificial Intelligence Research, Chinese Academy of Sciences
Pseudocode Yes Algorithm 1: Meta-training process
Open Source Code No The paper does not explicitly state that source code for the proposed Mecos framework is made publicly available.
Open Datasets Yes We use three public benchmark datasets for experiments. The first one is based on Steam (Kang and Mc Auley 2018), which contains user s reviews of online games from Steam, a large online video game platform. The second one is based on Amazon Electronic , which is crawled from amazon.com by (Mc Auley et al. 2015). The last one is based on Tmall from IJCAI-15 competition. This dataset contains user behavior logs in Tmall, the largest e-commerce platform in China. We apply the same preprocessing as (Kang and Mc Auley 2018; Tang and Wang 2018). Same as (Liu et al. 2018; Tan, Xu, and Liu 2016), the data augmentation strategy is employed on all datasets (e.g., a sequence (v0, v1, v2, v3) is divided into three successive sequences: (v0, v1), (v0, v1, v2), (v0, v1, v2, v3)). And that strategy is proved to be effective by previous studies (Liu et al. 2018; Tan, Xu, and Liu 2016).
Dataset Splits Yes Besides, the proportion of training, validation and testing tasks is 7 : 1 : 2. Moreover, we randomly leave out a subset of labels in Ytrain to generate the validation set Tmeta-valid.
Hardware Specification No The paper does not specify the hardware (e.g., CPU, GPU models, memory, or cluster configurations) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies (e.g., library names with version numbers) used for the implementation.
Experiment Setup Yes Each ground-truth next-click item is paired with 127 negative items (N = 128) randomly sampled from Ytest. All models are trained with the same data (pre-train data, all sequences in Tmeta-train and support sets in Tmeta-valid/Tmeta-test) for fair comparisons. Moreover, we vary the K to investigate the framework performance with different support set sizes and report K = 3 in default. Besides, the matching step t and hidden dimensionality d is set to 2 and 100, respectively. All hyper-parameters are chosen based on model s performances on Tmeta-valid. We run all models five times with different random seeds and report the average.